Player identification by motion features in sport videos using wearable sensors

Abstract

For real-time feedback and cost-efficient analysis from sport videos, it is essential to automatically identify players. In this paper, we propose a method for identifying sport players in videos. Our method uses wearable sensors to obtain their motions. Player identification is achieved by motion feature matching between (unknown) players in videos and wearable sensors whose IDs are already known. We combine three types of motion features, i.e. time sequences of speed, directions and step timings. For step detection from videos, we assume an existing computer vision technique to estimate postures (i.e. 18 joints of a skeleton) of players and design a step detection algorithm. Motion features from wearable sensors are extracted from acceleration, angular velocity and magnetic field. Simulation results show our method successfully identifies 10 players with 72 % accuracy at least even when average posture estimation error is 37.5 (cm).